You walk into a busy kitchen on a Friday night, the kind with an expediter standing at the pass calling out tickets faster than you can read them. A burger goes to the grill station without a second thought. A salad goes to the cold line. Fries get dropped by whoever is nearest the fryer. Nobody asks the head chef to personally plate a Caesar salad, and nobody expects the head chef to ignore the table that ordered a dish requiring an actual reduction sauce and twelve minutes of attention. The kitchen survives service not because everyone is equally skilled, but because someone upstream decided, ticket by ticket, who actually needs to touch it.
Now picture what happens when a company skips that decision and hands every single employee request to its most expensive, most capable AI model, regardless of whether the task is “summarize this email” or “restructure this quarter’s forecast.” That is most organizations’ current default, and it is quietly unworkable at any real scale. A frontier model is priced and rate-limited like a head chef, not a line cook, and a workforce of tens of thousands treating it like the latter turns a plausible pilot into a bill nobody signed off on, or a system so throttled it becomes useless right when everyone actually starts relying on it.
Cisco just ran into this problem at a scale most companies haven’t yet, and how they solved it is the concept worth understanding. In late August 2026, Cisco began rolling out a personal AI agent called MyAgent to essentially its entire 90,000-person workforce, one of the largest single-company agent deployments attempted so far. MyAgent sits on top of Cisco’s internal platform, Circuit, and is backed by more than 800 purpose-built subagents. An employee doesn’t pick a model or write a careful prompt. They state an objective, “reschedule this project’s dependencies” or “pull last quarter’s ticket volume by region,” and the system figures out which of those 800-plus subagents, and which underlying model, should actually handle it.
That’s the concept: intelligent routing beats uniform access. Cisco reports that roughly 50 to 60 percent of requests get handled by cheaper open-weight models, another 20 to 30 percent get resolved by plain software automation with no LLM call at all, and only a small remainder ever reaches a frontier model. The decision a leader has to make isn’t “which model should we buy,” it’s “who or what decides, per request, which model this task deserves,” and whether that decision is cheap enough and fast enough to run at company-wide volume.
The expensive model isn’t the product. The decision about when to use it is.
You’re probably thinking this sounds like an infrastructure problem, something for the platform team to sort out quietly, not a leadership call. I’d push back on that. The routing layer is also the governance layer: Cisco built a policy server into Circuit that blocks destructive actions and stops company data from training third-party models, and anything reaching outside the system still needs explicit human approval. That’s the control structure that makes it safe to let an agent touch Outlook, Webex, Jira, and SharePoint on 90,000 people’s behalf. In my experience, the organizations that get burned by agent rollouts aren’t the ones that picked the wrong model, they’re the ones that never decided who approves what before turning the system on.
| Uniform Model Access | Tiered Model Routing | |
|---|---|---|
| Default handler | Frontier model for every request | Cheapest capable option first |
| Cost at scale | Grows linearly with headcount | Bends down as routing absorbs volume |
| Governance | Bolted on per use case | Built into the routing/policy layer |
| Failure mode | Budget overrun or throttling | Misrouted task, needs an appeal path |
| Who decides | Individual employees, ad hoc | A defined routing + approval policy |
To be clear about what’s verified here: the 90,000-employee figure, the 800-plus subagent count, and the 50-60/20-30 routing split come from Cisco’s own disclosures, not an independent audit. Outside analysts have noted, fairly, that Cisco hasn’t published request volume, adoption curve, or an independent productivity measure, so nobody outside the company can say yet whether this saves money in absolute terms or merely avoids overspending relative to a worse design. It’s also worth noting the rollout landed in the same fiscal year as roughly 4,000 Cisco layoffs the company partly attributed to AI investment, a detail worth sitting with rather than skipping past. A cost-efficient routing system and a trusted rollout aren’t automatically the same achievement.
What doing this well concretely looks like, per Cisco’s own account: an employee’s request never touches a model directly. It goes to MyAgent, which classifies the objective, checks it against the policy server, and dispatches it to whichever of the 800-plus subagents (backed by the cheapest model capable of the job) can complete it, escalating to a frontier model only when the task’s complexity actually demands it. That routing decision, repeated at company-wide volume, is what keeps a 90,000-person deployment financially and operationally sane instead of a runaway line item.
If you’re rolling out AI access broadly this year, don’t start with which model to license. Start with this question for your team this week: for the ten most common AI requests employees will actually make, who or what decides which one gets the expensive model, and who signs off when that routing decision is wrong?
Sources: PYMNTS, Cisco Blogs, Fortune, UC Today